Liquid AI has introduced LFM2.5, a next-generation family of small foundation models built upon its innovative LFM2 architecture. These models are engineered with a specific focus on efficient deployment in on-device and edge computing environments. The diverse LFM2.5 series includes foundational and instruction-tuned variants, alongside specialized models for Japanese language, vision-language, and audio-language tasks. All models are openly available, with weights accessible on Hugging Face and integration facilitated through the LEAP platform.
LFM2.5: Engineered for Efficiency and Performance
The LFM2.5 models retain the hybrid LFM2 architecture, which is inherently optimized for rapid and memory-efficient inference on various processors, including CPUs and NPUs. A significant enhancement in this iteration is the scaling of the data and post-training pipelines. The pretraining phase for the 1.2 billion parameter core model was substantially expanded, progressing from 10 trillion to 28 trillion tokens. Furthermore, the instruct-tuned variant undergoes supervised fine-tuning, preference alignment, and extensive multi-stage reinforcement learning. This rigorous process concentrates on refining abilities such as instruction following, tool utilization, mathematical reasoning, and general knowledge inference.
Demonstrated Text Model Performance
The LFM2.5-1.2B-Instruct model serves as the primary general-purpose text solution within the family. Liquid AI has reported compelling benchmark results, positioning this model favorably against competitors in its class. On complex reasoning tasks, it achieved scores of 38.89 on GPQA and 44.35 on MMLU Pro. These figures notably surpass those of other 1-billion-parameter open models, including Llama-3.2-1B Instruct and Gemma-3-1B IT. For multi-step instruction adherence and function calling quality, the LFM2.5-1.2B-Instruct model posted strong results of 86.23 on IFEval and 47.33 on IFBench, again leading other comparable models.
Specialized Language and Multimodal Variants
Optimized for Japanese Language
A specialized version, LFM2.5-1.2B-JP, has been developed for the Japanese market, derived from the same robust backbone. This model is tailored for Japanese-specific tasks like JMMLU, M-IFEval in Japanese, and GSM8K in Japanese. It demonstrates superior performance on localized benchmarks compared to the general instruct model and either matches or exceeds other small multilingual models such as Qwen3-1.7B, Llama 3.2-1B Instruct, and Gemma 3-1B IT.
Vision-Language for Real-World Edge Applications
The updated LFM2.5-VL-1.6B model brings enhanced vision-language capabilities to the series. It integrates a dedicated vision tower for image comprehension with the LFM2.5-1.2B-Base as its language foundation. This model has undergone tuning across a broad spectrum of visual reasoning and optical character recognition (OCR) benchmarks, including MMStar, MM IFEval, and OCRBench v2, among others. It shows significant improvements over its predecessor, LFM2-VL-1.6B, and is designed for practical applications such as document analysis, user interface interpretation, and multi-image reasoning in constrained edge environments.
Native Audio Interaction and Generation
LFM2.5-Audio-1.5B stands out as a native audio language model, supporting both audio and text inputs and outputs. Presented as an Audio-to-Audio model, it incorporates an audio detokenizer reportedly eight times faster than previous versions while maintaining precision on hardware with limited resources. The model offers two distinct generation modes: interleaved generation for real-time speech-to-speech conversational agents where low latency is critical, and sequential generation for tasks such as automatic speech recognition (ASR) and text-to-speech (TTS), allowing modality switching without reinitialization. Its audio stack benefits from quantization-aware training at low precision, ensuring performance metrics remain consistent with full-precision baselines, thereby facilitating deployment on devices with restricted computational power.
Availability and Impact
The release of the LFM2.5 family as open weights on Hugging Face and through the LEAP platform signifies Liquid AI's commitment to advancing accessible, high-performance AI for on-device applications. By offering a range of specialized models—from general text to multimodal and localized variants—the LFM2.5 series is poised to empower developers and organizations to create sophisticated AI agents capable of operating effectively at the edge.
This article is a rewritten summary based on publicly available reporting. For the original story, visit the source.
Source: MarkTechPost
